SCCM : Stream Cruise Control Method for Automated Drift Detection and Adaptation

Motivation

• Online learning is needed for non-stationary environments because incremental updates enable real-time responses and continuous adaptation, unlike conventional batch-trained models.Concept drift arises as data distributions evolve, making continual manual hyperparameter tuning impractical.Source: §1 · p. 2

Limitations

• Existing methods often detect drift reactively after performance degradation, provide only binary alarms, and separate detection from proportionate adaptation.Ensemble approaches can improve robustness but may impose substantial computational overhead in high-velocity or resource-constrained streams.Source: §1 · p. 3

SCCM framework

• SCCM operates as a pre-update control layer that monitors predictive performance, detects drift, quantifies severity, and adjusts model hyperparameters before learner updates.It unifies detection, magnitude estimation, and adaptive hyperparameter control rather than replacing the underlying learner.Source: §1 · pp. 3–4

Algorithmic mechanism

• SCCM uses bounded KPI-window memory, transient pre-update observations, CFAR-inspired local thresholds, severity regions, scale-mapped hyperparameter control, and conditional bounded recalibration.The framework distinguishes stable, incremental, and abrupt drift and maps severity to model-specific settings such as forgetting factors or regularization parameters.Source: §1 · pp. 4–5

Adaptation and deployment

• SCCM adaptively controls learner hyperparameters after deployment-time configuration, avoiding repeated offline tuning while supporting model-agnostic, bounded-memory, constant-size streaming deployment.Configuration parameters such as z and the selected KPI remain user-defined, while learning-dynamic settings are adjusted online.Source: §1 · pp. 5–6

Evaluation scope

• The study evaluates SCCM with multiple online regressors under abrupt, incremental, and alternating gradual drift, comparing it with detectors, adaptation strategies, and online hyperparameter-tuning methods.Evaluation considers predictive performance and drift-handling effectiveness across varying dataset characteristics.Source: §1 · p. 6

2. Related Work

• The literature review positions SCCM against detectors and adaptation baselines including ADWIN, KSWIN, RESET, WINDOW, OHL, and SSPT, emphasizing regression’s relative underrepresentation.Tables 1–3 summarize detection, adaptation, and framework methods, while the study uses the listed methods as experimental baselines.Source: §2 · p. 7, §2.5 · p. 16

2.1. Concept Drift Detectors

• Drift detectors are categorized as distribution-oriented or performance-oriented, but classification-focused methods, detector sensitivity, delays, and computational overhead limit generality for online regression.ADWIN compares statistics across adaptive subwindows but may incur delay and memory overhead, whereas KSWIN is distribution-free but sensitive to window settings and less suitable for high-dimensional data.Source: §2.1 · pp. 7–8, §2.5 · p. 16

2.2.1. Hyperparameters Optimization

• Batch hyperparameter optimization relies on repeated training and validation over fixed data, conflicting with single-pass online learning and limited access to past observations.Online alternatives include OHL’s projected hyper-gradient updates [28], SPOT, SSPT’s single-pass Nelder–Mead search, Hyperband for Streams, and HyperTube.Source: §2.2.1 · pp. 10–11

2.2.2. Adaptation Strategies

• Adaptation strategies include model replacement, recent-window retraining, ensembles, and structural updates, trading off simplicity, knowledge retention, computational cost, and responsiveness.RESET reinitializes the model, WINDOW uses recent observations, ensembles combine learners but can be expensive, and methods such as CVFDT and DWM modify model structure.Source: §2.2.2 · p. 11, §2.5 · p. 16

2.3. Automated Drift Adaptation Frameworks

• Automated adaptation frameworks integrate detection, adaptation, and maintenance, but reviewed examples remain domain-specific, resource-intensive, history-dependent, or tied to particular model families.StepWise targets web anomaly detection, AEF-CDA targets large-scale healthcare streams, and HyperTube targets constrained deep-learning optimization using stored history.Source: §2.3 · pp. 12–13, §2.5 · p. 16

2.4. Adaptive Hyperparameters in Online Regression

• Online regression hyperparameters regulate historical-versus-recent information through forgetting, weighting, aggressiveness, learning-rate, and momentum controls, yet adapting them under drift remains difficult.Examples include RLS’s forgetting factor λ, OLR-WA’s weighting parameter α, PA’s aggressiveness parameter C, LMS’s learning rate, and SGDM’s momentum coefficient β.Source: §2.4 · pp. 13–14
• SCCM addresses a gap in unified regression-oriented drift handling by combining KPI-based detection, magnitude estimation, proportional hyperparameter tuning, and bounded recalibration in one pre-update process.Existing methods typically address only detection, retraining, replacement, or hyperparameter tuning, while RESET, WINDOW, OHL, and SSPT require an external drift signal.Source: §2.4 · p. 14

2.5. Summary

• Overall, existing methods provide complementary capabilities but lack a general-purpose, low-latency framework that jointly detects drift, controls hyperparameters, and adapts online for regression.The review highlights distinct trade-offs across drift types, latency, computational cost, and knowledge retention.Source: §2.4 · p. 14, §2.5 · p. 16

• Online learning processes an effectively infinite data stream sequentially while retaining only limited data in memory, and concept drift denotes unexpected distributional change over time.Data points are typically handled in small increments.Source: §3 · p. 16
• Concept drift duration is the interval between an initial concept a_n ≡ p_tn(X, y) and resulting concept a_n+1 ≡ p_tn+1(X, y), measured as t_n+1 − t_n.The stream is modeled by a time-varying joint density p(X, y).Source: §3 · pp. 16–17
• Drift magnitude is the distribution distance D(a_0, a_1) between concepts over a drift period, with minor drift favoring refinement and major drift potentially requiring model abandonment.Magnitude and duration influence learner selection and adaptation.Source: §3 · pp. 16–17
• Drift may be abrupt, incremental, or gradual, and may be global across the distribution or local to a subset or feature-space region.Local drift can leave overall stream behavior relatively stable while changing part of the input-output relationship.Source: §3 · p. 17
• Unknown drift duration and magnitude, together with data noise, make conventional detection intractable and motivate an easily implemented method applicable to diverse drift types.Effective drift-aware online learning depends heavily on reliable drift detection.Source: §3 · p. 17

Framework operation

• SCCM evaluates each incoming instance or mini-batch before weight updates, retaining it temporarily and triggering corrective adaptation only when KPI behavior indicates drift.Without detected drift, standard online updating proceeds; detected drift initiates characterization and adaptation.Source: §4 · p. 18

KPI monitoring

• Recent KPI values are stored in a bounded KPI-Window with historical baseline entries and a transient InstKPI entry for point- or mini-batch-level monitoring.Regression KPIs can include MAE, MSE, RMSE, and R2; the window size is bounded using recent observations, mini-batch size, and γ = 0.05 in the experiments.Source: §4 · p. 19

Thresholding and severity

• SCCM computes KPI-dependent dynamic thresholds from recent-window statistics and classifies observations as stable, incremental drift, or abrupt drift using a safe band and outer limits.The safe-band tolerance ζ controls sensitivity; ζ = 0.005 was used experimentally to balance responsiveness and stability.Source: §4 · pp. 20–21

Local false-alarm control

• The user-controlled nominal level ρ calibrates local detection sensitivity through z, but it provides no global stream-level false-alarm guarantee under repeated non-stationary decisions.Higher ρ lowers z and increases sensitivity, whereas lower ρ raises z and produces more conservative thresholds; the safe band absorbs benign fluctuations.Source: §4 · pp. 24–25

Adaptive control and recalibration

• Detected drift activates a deviation-scaled hyperparameter map, assigning stronger adaptation to larger deviations while preserving stability near the safe-band boundary.If drift persists beyond the outer limits after optimization, SCCM performs bounded recalibration using additional incoming instances, typically 3 to 5 in experiments.Source: §4 · pp. 25–26

Algorithmic distinction

• SCCM unifies KPI-based memory, local CFAR-inspired thresholding, drift-magnitude estimation, severity classification, proportional hyperparameter control, and bounded recalibration in one pre-update pipeline.Unlike methods requiring external drift triggers, SCCM converts recent KPI deviations directly into adaptive control for online regression while remaining extensible beyond regression.Source: §4 · pp. 18, 26, 27

Experimental scope

• SCCM is evaluated under diverse non-stationary environments through a comprehensive experimental study.The experiments are designed to assess the proposed framework across changing data conditions.Source: §5 · p. 27

Baseline comparison

• Table 5 compares SCCM with representative concept drift detection and adaptation methods at the algorithmic level.The supplied passage identifies the comparison scope but does not provide quantitative cell values.Source: Table 5 · p. 28

Datasets

• The evaluation uses both synthetic and real-world datasets to assess SCCM’s robustness, generality, and practical applicability.These dataset types provide complementary settings for evaluating the proposed approach.Source: §5 · p. 29

5.1.1. Synthetic Datasets

• Synthetic datasets model abrupt, incremental, and alternating gradual drift by varying Gaussian input distributions, linear-regression parameters, dimensionality, drift magnitude, and sample size.Drift is induced through changes in input means, regression coefficients, and bias terms; target noise has standard deviation 1.5.Source: §5.1.1 · pp. 29–30
• The synthetic benchmark spans mild to strong distributional changes, including target Wasserstein distances from 1.2103 to 7.983 and significant feature-level drift across all datasets.Examples include ADS04’s mild abrupt drift and ADS05’s stronger abrupt shift, while GDS04 is mild and GDS02/GDS05 show stronger alternating gradual drift.Source: §5.1.1 · p. 30

5.1.2. Real-World Datasets

• Eight real-world regression datasets span healthcare, business and finance, housing and real estate, energy systems, environmental monitoring, chemical sensing, oceanography, and retail sales forecasting.Source: §5.1.2 · p. 30
• Real-world streams are analyzed chronologically by comparing their first and second halves, treating drift measures as empirical non-stationarity indicators rather than controlled magnitudes.The analysis jointly considers target and feature changes, fitted-parameter distance, KS tests, Jensen–Shannon divergence, and Wasserstein distance.Source: §5.1.2 · pp. 30–33

5.2. Experimental Setup

• Synthetic datasets cover abrupt, incremental, and alternating gradual drift while controlling drift magnitude, feature dimensionality, and noise level.This design supports reproducible analysis of model behavior under different distributional changes.Source: §5.2 · p. 35
• Four online regressors—RLS, PA, LMS, and OLR-WA —are evaluated before and after SCCM integration.The models represent diverse update dynamics, from reactive updates to stable incremental adaptation.Source: §5.2 · p. 35
• Each model is compared with eight detector–adaptation baselines combining ADWIN or KSWIN with RESET, WINDOW, SSPT, or OHL.RESET reinitializes models, WINDOW retrains on recent data, SSPT searches hyperparameters with Nelder–Mead, and OHL updates hyperparameters using online hyper-gradients.Source: §5.2 · pp. 35–36

5.2.1. Baseline Configuration Protocol

• Baseline detector and adaptation settings remain fixed across models, datasets, and seeds, with no offline grid search or future-stream tuning.ADWIN uses δ = 0.002; KSWIN uses αKS = 0.005, WKS = 100, and SKS = 30.2, while configurations are unchanged across five seeds.Source: §5.2.1 · pp. 36–37
• Results are aggregated by model, method, and drift type across datasets and five seeds, using R2 for mini-batch assessment and MSE for online and mini-batch evaluation.Alarm metrics are pooled at the seed level and reported as mean ± standard deviation over five seeds.Source: §5.2.1 · pp. 37, 39

5.3.1. Predictive Performance Metrics

• R2 evaluates mini-batch models such as OLR-WA, whereas MSE evaluates fully online PA, RLS, and LMS models using sequential point-wise predictions.R2 is normalized and scale-invariant, while MSE preserves absolute error magnitude and penalizes larger deviations more strongly.Source: §5.3.1 · pp. 39–40

5.3.2. Predictive Statistical Protocol

• Performance is summarized across five seeds and compared with standalone learners and all eight detector–adaptation baselines using paired two-sided Wilcoxon tests with Holm correction.The strongest baseline is identified for presentation, but inference uses all eight comparisons; rank-biserial correlation reports effect size, with positive MSE effects favoring lower SCCM error.Source: §5.3.2 · p. 40

5.3.3. Drift-Alarm Quality Metrics and Alignment Protocol

• Synthetic streams report alarm precision, recall, F1-score, false positives, missed drifts, and detection delay, while real-world streams lack ground-truth alarm evaluation.Primary analysis retains every raw ADWIN/KSWIN alarm and matches alarms to known drifts chronologically and one-to-one; episode consolidation is secondary, and episode-level F1 is the primary inferential endpoint.Source: §5.3.3 · pp. 40, 41, 43

5.4.1. Performance Analysis on Abrupt Datasets

• Under abrupt drift, SCCM-integrated OLR-WA, PA, and RLS variants improve average error or fit over standalone models, with OLR-WA* achieving the strongest OLR-WA-based performance.OLR-WA* reaches average R2 0.804 versus 0.715 and average MSE 1.724 versus 2.157; PA* reaches average MSE 1.246 versus 5.323; RLS* reaches 0.908 versus 3.906.Source: §5.4.1 · pp. 43, 45

5.4.2. Performance Analysis on Incremental Datasets

• Under incremental drift, SCCM-integrated PA, RLS, and LMS variants achieve lower average MSE than PA-KS or standalone counterparts, recording 1.424 versus 4.469, 0.813 versus 2.698, and 0.536 versus 3.522.Source: §5.4.2 · p. 51

5.4.3. Performance Analysis on Alternating Gradual Datasets

• Under alternating gradual drift, SCCM-integrated OLR-WA, PA, and RLS variants achieve the best performance within their respective model families, with lower errors during transitions between old and new concepts.OLR-WA∗ achieves average R2 0.797 versus 0.622 and MSE 1.855 versus 3.093; PA∗ achieves MSE 1.532 versus 7.034; and RLS∗ achieves MSE 1.227 versus 7.163.Source: §5.4.3 · pp. 54, 57

5.4.4. Predictive Uncertainty and Statistical Significance

• Across all four models and three synthetic drift categories, SCCM improves performance over the strongest detector–adaptation baselines, reduces variability, and remains statistically significant after correction.Mean MSE reductions are 68.1%–78.1% for PA, 69.7%–78.2% for RLS, and 84.7%–88.3% for LMS; adjusted p-values are no larger than 3.7×10−3, with rank-biserial correlations from +0.58 to +0.73.Source: §5.4.4 · pp. 60, 62, Table 10 · p. 60

5.4.5. Performance Analysis on Real Datasets

• On eight real-world chronological datasets, the evaluation assesses predictive robustness under natural temporal variation rather than drift-detection accuracy because drift locations, types, and boundaries are unavailable.Results are aggregated over five independent random seeds, using R2 for OLR-WA and MSE for PA, RLS, and LMS.Source: §5.4.5 · p. 62
• The evaluation compared standalone learners, SCCM variants, and eight detector–adaptation baselines across four models and eight real-world datasets using five evaluation seeds.Higher R2 indicates better performance for OLR-WA, whereas lower MSE indicates better performance for PA, RLS, and LMS.Source: §5.4.5 · pp. 62–63
• Comparisons with the strongest observed detector–adaptation baselines were descriptive post hoc analyses because those baselines were selected from the same reported real-world results.Confirmatory Wilcoxon signed-rank tests used eight dataset-level paired observations and Holm correction across four standalone-learner comparisons.Source: §5.4.5 · p. 63
• Median gains over standalone learners were 0.0586 R2 for OLR-WA and 76.4%–80.0% MSE reductions for PA, RLS, and LMS.All Holm-adjusted p-values were below 10^-6, with rank-biserial correlations from +0.944 to +1.000.Source: §5.4.5 · pp. 63, 66, Table 13 · p. 66
• SCCM-integrated variants achieved the best mean predictive performance across all 32 real-world model–dataset combinations, improving R2 for OLR-WA and reducing MSE for PA, RLS, and LMS.Gains spanned datasets with different domains, sample sizes, dimensionalities, and temporal variation, with generally lower five-seed variability.Source: Table 11 · p. 64, Table 12 · p. 65, §5.4.5 · p. 66

5.5. Drift Detection and Adaptation-Trigger Analysis

• SCCM adaptation and recalibration counts represent pre-consolidation internal trigger activity and are not equivalent to retained alarm episodes, false-alarm counts, or direct computational-cost measurements.Unlike external detector pipelines, SCCM may perform multiple proportional adjustments and recalibrations within one continuing alarm episode.Source: §5.5 · p. 67, §5.5.1 · p. 68, §5.5.4 · p. 75

5.5.2. Illustrative Alarm-Episode Construction and Matching

• Episode-level matching uses each SCCM episode’s first trigger, so pre-drift episodes remain false positives despite later post-drift triggers; retained episodes are matched chronologically one-to-one with known drifts.Source: §5.5.1 · p. 69, §5.5.2 · p. 71

5.5.1. Direct Raw-Alarm Evaluation of ADWIN and KSWIN

• Raw ADWIN and KSWIN alarms are evaluated individually, yielding identical detector-level sequences across RESET, WINDOW, SSPT, and OHL, while KSWIN generally outperforms ADWIN in alarm quality.KSWIN’s raw F1-scores are 0.325 ± 0.022 for abrupt and 0.643 ± 0.019 for alternating gradual drift, compared with ADWIN’s 0.278 ± 0.018 and 0.616 ± 0.020; incremental recall remains low for both.Source: §5.5.1 · pp. 67–69, Table 14 · p. 69

5.5.3. SCCM Episode-Level Alarm-Quality Analysis

• SCCM-integrated learners maintain recall above 0.86 in every model–drift combination, while incremental and alternating gradual drift are handled consistently better than abrupt drift.For incremental drift, PA* attains the highest F1-score of 0.9486; for alternating gradual drift, PA* and LMS* reach 0.9701, while abrupt drift exposes more unmatched episodes for PA*, RLS*, and LMS*.Source: §5.5.3 · pp. 71, 73

5.5.4. Episode-Level Detector–Adaptation Benchmark

• SCCM detects matched drifts faster than ADWIN and KSWIN across all categories, with mean delays of 0.29, 3.20, and 1.07 processing increments for abrupt, incremental, and alternating gradual drift.Under abrupt drift, SCCM’s delay is 0.29 increments versus 5.68 for ADWIN and 5.27 for KSWIN; delay must be interpreted jointly with recall.Source: §5.5.4 · p. 75
• SCCM achieves higher episode-level F1-scores than ADWIN and KSWIN under abrupt, incremental, and alternating gradual drift, with the largest advantages under incremental and alternating gradual drift.SCCM’s F1-scores are 0.7110, 0.9296, and 0.9676, compared with strongest detector baselines of 0.5113, 0.1431, and 0.2032, respectively.Source: Table 16 · p. 74, §5.5.4 · p. 75

5.5.5. Paired Statistical Comparison of Alarm Quality

• Across 120 paired model–dataset–seed observations per comparison, SCCM’s episode-level F1-score is significantly higher than both detector baselines in every drift category after Holm correction.All six comparisons remain significant, with large effects under abrupt drift and near-uniform advantages under incremental and alternating gradual drift.Source: §5.5.5 · pp. 76–77, Table 17 · p. 77

5.5.6. Sensitivity to Alarm-Evaluation Parameters

• The sensitivity analysis independently varies rtol, ccool, and mep while fixing the other parameters at the primary setting (0.05, 2.0, 2).Table 18 reports SCCM precision, recall, F1, and detection delay, while ADWIN and KSWIN are summarized by F1 and detection delay; detector-family configurations produce identical alarm sequences under matched execution settings.Source: §5.5.6 · p. 78
• Changing rtol yields SCCM F1-scores from 0.5779 at rtol = 0.025 to 0.7686 at rtol = 0.100 under abrupt drift, while rtol = 0.050 gives 0.9296 under incremental and 0.9676 under alternating gradual drift.Cooldown changes have comparatively small effects; increasing mep generally raises precision and lowers recall, with mep = 2 best under incremental and alternating gradual drift and effectively tied with mep = 3 under abrupt drift.Source: §5.5.6 · p. 78
• Across every evaluated parameter setting, SCCM achieves higher F1-scores and shorter detection delays than ADWIN and KSWIN.Sensitivity is greatest under abrupt drift and stability is greatest under alternating gradual drift.Source: §5.5.6 · p. 78

5.6. Computational Cost and Normalized Intervention Activity

• Runtime and memory were measured single-threaded in the same Lonestar6 environment, with runtime normalized per 1,000 samples and peak memory recorded for learners and drift-handling variants.The environment used two AMD EPYC 7763 64-core processors, approximately 256 GB memory, Rocky Linux 8.10, and Python 3.9.7.Source: §5.6 · pp. 78, 80
• SCCM requires 5.68 ± 0.44 ms per 1,000 samples and 4.35±0.33 MB peak memory, remaining below every evaluated WINDOW, SSPT, and OHL configuration.This corresponds to a 52.7% runtime overhead relative to the standalone learner; RESET is slightly faster at 4.91 ± 0.39 ms with ADWIN and 5.37±0.42 ms with KSWIN, while SSPT reaches 13.46 ± 0.91 ms and 14.12 ± 0.96 ms.Source: §5.6 · pp. 80–81, Table 19 · p. 81
• SCCM records 452.80 adaptations and 361.81 recalibrations per 1,000 samples, versus 2.30 detector-triggered adaptations for ADWIN and 2.38 for KSWIN.SCCM adaptations are bounded model-specific hyperparameter assignments, whereas baseline events may involve resetting, recent-window retraining, or iterative hyperparameter optimization, so frequencies are not equivalent operations.Source: §5.6 · pp. 80–81
• Overall, SCCM is more demanding than the standalone learner and RESET but substantially more efficient than WINDOW, SSPT, and OHL configurations.Its lightweight characterization refers to bounded memory and inexpensive control operations, not zero overhead or few intervention signals.Source: §5.6 · pp. 80–81

Ablation design

• The analysis evaluates bounded recalibration, the safe band ζ, nominal sensitivity ρ, and KPI-window size against the base model and full SCCM under concept drift.Variants include SCCM without bounded recalibration or the safe band, alongside different ρ and KPI-window settings; Table 20 summarizes these comparisons.Source: §5.7 · pp. 81–82, Table 20 · p. 82

Experimental setting

• The study uses OLR-WA on representative synthetic streams with abrupt and alternating gradual drift to expose different adaptation requirements.OLR-WA’s adaptive weighting parameter α makes SCCM’s hyperparameter-control mechanism easy to interpret; abrupt drift requires fast correction, whereas alternating gradual drift requires stable adaptation.Source: §5.7 · pp. 81–82

Sensitivity to ρ

• A smaller ρ increases z and makes SCCM more conservative, whereas a larger ρ narrows the acceptable region, improving subtle-drift sensitivity but potentially increasing unnecessary adaptations under noise.The threshold follows z = Φ−1(1 −ρ).Source: §5.7 · pp. 82–83

KPI-window sensitivity

• Smaller KPI windows improve responsiveness but can destabilize local estimates, while larger windows stabilize baselines at the cost of delayed detection.The window size therefore trades off responsiveness against stability.Source: §5.7 · p. 83

Overall findings

• The complete SCCM framework provides the best balance between responsiveness and stability, integrating monitoring, sensitivity control, fluctuation suppression, proportional tuning, and bounded recalibration.KPI-window monitoring establishes the local baseline; ρ controls detection sensitivity; the safe band suppresses minor fluctuations; drift magnitude drives proportional hyperparameter adjustment; and bounded recalibration addresses persistent severe drift.Source: §5.7 · p. 83

5.8. Configuration and Hyperparameter Settings

• The experimental configurations and hyperparameters are reported to clarify, reproduce, and consistently apply the evaluated models and baselines.Source: §5.8 · p. 83

5.8.1. OLR-WA-Based Methods

• OLR-WA and its drift-adaptive variants use configuration settings and definitions summarized in Table 21.Source: §5.8.1 · p. 83, Table 21 · p. 84

5.8.2. PA-Based Methods

• The PA model and its drift-adaptive variants use configuration settings and definitions summarized in Table 22.Source: §5.8.2 · p. 83, Table 22 · p. 85

5.8.3. RLS-Based Methods

• The RLS model and its drift-adaptive variants use configuration settings and definitions summarized in Table 23.Source: §5.8.3 · p. 83, Table 23 · p. 86

5.8.4. LMS-Based Methods

• The LMS model and its drift-adaptive variants use configuration settings and definitions summarized in Table 24.Source: §5.8.4 · p. 83, Table 24 · p. 87

Conclusion

• SCCM is a model-agnostic pre-update control layer that monitors predictive behavior, detects and quantifies drift, tunes model-specific hyperparameters, and performs bounded recalibration.Its unified mechanism combines KPI-window monitoring, CFAR-inspired adaptive thresholding, severity-aware classification, and scale-map-based hyperparameter control.Source: §6 · p. 88
• Across 18 synthetic and eight real-world datasets, SCCM improves overall predictive performance when integrated with four online regression models and compared with eight detector–adaptation baselines.The evaluation uses R2 and MSE across abrupt, incremental, and alternating gradual drift and heterogeneous real-world domains, scales, and dimensionalities.Source: §6 · p. 88
• Synthetic-stream analysis evaluates detection alarms, true and false alarms, missed drifts, precision, recall, F1-score, delay, protocol sensitivity, runtime, memory, and intervention activity.Direct drift-detection accuracy is validated only on synthetic streams because ground-truth drift locations are unavailable for the real-world datasets.Source: §6 · p. 88
• Real-world experiments assess predictive robustness under naturally occurring, unlabeled temporal variation, supporting SCCM as a practical and extensible framework within the evaluated online regression settings.The conclusion distinguishes real-world predictive evaluation from controlled synthetic drift-detection validation.Source: §6 · p. 88

GenAI Usage Disclosure

• Generative AI supported language refinement exclusively, while the authors independently conducted the method design, implementation, analysis, and evaluation.The disclosure states that AI use did not include content generation that could compromise scientific integrity and complies with Elsevier’s Authorship Policy.Source: GenAI Usage Disclosure · p. 88